# Agentic workflow

An agentic workflow is multi-step work where the sequence is decided as it goes rather than written in advance. A script knows its next step; an agentic workflow knows its goal and picks the step. That flexibility is the whole benefit and the whole risk, which is why the interesting design question is where the checkpoints go.

- **Status:** Reference
- **Audience:** both, developer
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/technology/agentic-workflow/

## A worked example: one prospect, eleven steps

A business asks Connect to find suppliers of a component in a named region and open a conversation with the ones worth talking to. Nothing about that request names a sequence. What happens is:

1. Read the discovery brief and turn it into search criteria.
2. Collect candidate organisations from public sources.
3. Discard the ones already in the workspace as customers or prospects.
4. Research each survivor against the brief, cheaply.
5. Score fit, and record the evidence for the score.
6. Spend a deeper research pass only where it could change the decision.
7. Look for a real, published contact address — never a constructed one.
8. Check suppression and do-not-contact across every channel.
9. Draft an opening that cites the specific reason this organisation was chosen.
10. Hold or send, according to the channel's autonomy setting.
11. Record the outcome, the cost, and the next action.

Steps four to seven vary per candidate: one organisation is resolved in a single pass, another needs the deep one, a third is dropped at step three. That variability is what makes it agentic. Steps eight and ten are fixed and are not the model's to skip.

## Where a checkpoint earns its place

| Put a checkpoint | Because | Do not put one |
|---|---|---|
| Before the first irreversible act | A sent message cannot be unsent, and this is the last moment a person can change the outcome. | Before each read, which teaches people to click through. |
| Where the rules already reserve a decision | A price, an exception, a commitment — the business has said these are not the agent's. | On low-risk work the business explicitly delegated. |
| Where cost steps up | A deep research pass, a paid model call, an outbound call. | On every model call, which makes the queue useless. |
| Where confidence is genuinely low | An unresolved identity, a conflicting source, a promise with no time in it. | Everywhere, as a substitute for measuring confidence. |

> **Careful** A checkpoint that fires on everything is not a control. It is a training exercise in approving without reading, and the first genuinely wrong draft goes out with the rest.

## Does Connect use agentic workflows?

**Used.** Prospecting, mail triage and reply, call handling and follow-up execution are all of this shape, and they share one thirteen-stage skeleton described at [how Connect works](/docs/product/how-connect-works/). The variability lives in the classification, research and reasoning stages; the permission, action and audit stages are fixed.

The checkpoint is a single named stage — autonomy and approval — rather than a decision scattered through the flow. That is deliberate: a checkpoint that can appear anywhere is a checkpoint nobody can enumerate, and enumerating them is how [Needs You](/docs/autonomy/needs-you/) can rank them.

## Failure modes worth knowing before you design one

- **The loop that will not end.** Without a terminating condition outside the model, an agent re-plans the same failing step. Connect bounds work by budget, by allowance and by explicit retry rules rather than by asking the model to give up.
- **The half-done sequence.** Step seven fails and steps one to six have already changed things. Every stage here records what it completed, so a recovery starts from the record rather than from the beginning.
- **The confident wrong branch.** The model picks a plausible plan for the wrong goal. Grounding and evidence reduce this; a checkpoint before the first irreversible act contains what is left.
- **Silent success.** The workflow completes and nobody can say what it did. An agentic flow without an audit entry per stage is unauditable by construction.

## Reading further

The nearest thing to a specification for describing multi-step API work is the OpenAPI Initiative's Arazzo, version 1.1.0 ([spec.openapis.org/arazzo](https://spec.openapis.org/arazzo/latest.html)). It describes deterministic sequences rather than agentic ones, which is exactly the contrast this page is about — and Connect publishes its own workflows in that format for machine readers. See [Arazzo](/docs/protocols/arazzo/).

## Questions

### How is this different from a workflow automation tool?

An automation tool executes a sequence somebody drew. An agentic workflow chooses the sequence against a goal. The automation tool is more predictable and cannot handle the case its author did not foresee; the agent handles it and is harder to predict. Neither is better in general.

### Can I see the steps an agentic workflow took?

In Connect, yes — the audit trail records the decision at each stage, the rule that applied and the result, which is what makes 'why did it do that?' answerable months later.

### What happens if a step fails halfway through?

The work stops at that stage with a recorded reason, and an item appears in Needs You if a person can act on it. Nothing continues silently past a failure and nothing is retried in a way that could send twice.

## Related

- [AI agent](https://connectbyjbrh.com/docs/technology/ai-agent/)
- [Agent orchestration](https://connectbyjbrh.com/docs/technology/orchestration/)
- [End-to-end workflows](https://connectbyjbrh.com/docs/workflows/)
- [Approvals](https://connectbyjbrh.com/docs/autonomy/approvals/)
- [Arazzo](https://connectbyjbrh.com/docs/protocols/arazzo/)

## What this page is based on

- AGENTS.md §3 — how a request flows
- Connect capability registry (docs-source/facts.py) — prospect_* capabilities
- Arazzo Specification 1.1.0 (https://spec.openapis.org/arazzo/latest.html)